ICML 2024 Papers

Pre-trained vision-language models, e.g., CLIP, have been successfully applied to zero-shot semantic segmentation. Existing CLIP-based approaches primarily utilize visual features from the last layer to align with text embeddings, while they neglect the crucial information in intermediate layers that contain rich object details. However, we find that directly aggregating the multi-level visual features weakens the zero-shot ability for novel classes. The large differences between the visual features from different layers make these features hard to align well with the text embeddings. We resolve this problem by introducing a series of independent decoders to align the multi-level visual features with the text embeddings in a cascaded way, forming a novel but simple framework named Cascade-CLIP. Our Cascade-CLIP is flexible and can be easily applied to existing zero-shot semantic segmentation methods. Experimental results show that our simple Cascade-CLIP achieves superior zero-shot performance on segmentation benchmarks, like COCO-Stuff, Pascal-VOC, and Pascal-Context. Our code is available at https://github.com/HVision-NKU/Cascade-CLIP.

We study the data selection problem, whose aim is to select a small representative subset of data that can be used to efficiently train a machine learning model. We present a new data selection approach based on $k$-means clustering and sensitivity sampling. Assuming access to an embedding representation of the data with respect to which the model loss is Holder continuous, our approach provably allows selecting a set of ``typical'' $k + 1/\varepsilon^2$ elements whose average loss corresponds to the average loss of the whole dataset, up to a multiplicative $(1\pm\varepsilon)$ factor and an additive $\varepsilon \lambda \Phi_k$, where $\Phi_k$ represents the $k$-means cost for the input embeddings and $\lambda$ is the Holder constant. We furthermore demonstrate the performance and scalability of our approach on fine-tuning foundation models and show that it outperforms state-of-the-art methods. We also show how it can be applied on linear regression, leading to a new sampling strategy that surprisingly matches the performance of leverage score sampling, while being conceptually simpler and more scalable.

Significant progress has been achieved on the improvement and downstream usages of the Contrastive Language-Image Pre-training (CLIP) vision-language model, while less attention is paid to the interpretation of CLIP. We propose a Gradient-based visual Explanation method for CLIP (Grad-ECLIP), which interprets the matching result of CLIP for specific input image-text pair. By decomposing the architecture of the encoder and discovering the relationship between the matching similarity and intermediate spatial features, Grad-ECLIP produces effective heat maps that show the influence of image regions or words on the CLIP results. Different from the previous Transformer interpretation methods that focus on the utilization of self-attention maps, which are typically extremely sparse in CLIP, we produce high-quality visual explanations by applying channel and spatial weights on token features. Qualitative and quantitative evaluations verify the superiority of Grad-ECLIP compared with the state-of-the-art methods. A series of analysis are conducted based on our visual explanation results, from which we explore the working mechanism of image-text matching, and the strengths and limitations in attribution identification of CLIP. Codes are available here: https://github.com/Cyang-Zhao/Grad-Eclip.

The increasing complexity of foundational models underscores the necessity for explainability, particularly for fine-tuning, the most widely used training method for adapting models to downstream tasks. Instance attribution, one type of explanation, attributes the model prediction to each training example by an instance score. However, the robustness of instance scores, specifically towards dataset resampling, has been overlooked. To bridge this gap, we propose a notion of robustness on the sign of the instance score. We theoretically and empirically demonstrate that the popular leave-one-out-based methods lack robustness, while the Shapley value behaves significantly better, but at a higher computational cost. Accordingly, we introduce an efficient fine-tuning-free approximation of the Shapley value (FreeShap) for instance attribution based on the neural tangent kernel. We empirically demonstrate that FreeShap outperforms other methods for instance attribution and other data-centric applications such as data removal, data selection, and wrong label detection, and further generalize our scale to large language models (LLMs). Our code is available at https://github.com/JTWang2000/FreeShap.

We present the Hourglass Diffusion Transformer (HDiT), an image-generative model that exhibits linear scaling with pixel count, supporting training at high resolution (e.g. $1024 \times 1024$) directly in pixel-space. Building on the Transformer architecture, which is known to scale to billions of parameters, it bridges the gap between the efficiency of convolutional U-Nets and the scalability of Transformers. HDiT trains successfully without typical high-resolution training techniques such as multiscale architectures, latent autoencoders or self-conditioning. We demonstrate that HDiT performs competitively with existing models on ImageNet $256^2$, and sets a new state-of-the-art for diffusion models on FFHQ-$1024^2$. Code is available at https://github.com/crowsonkb/k-diffusion.

Training and validating models for semantic segmentation require datasets with pixel-wise annotations, which are notoriously labor-intensive. Although useful priors such as foundation models or crowdsourced datasets are available, they are error-prone. We hence propose an effective framework of active label correction (ALC) based on a design of correction query to rectify pseudo labels of pixels, which in turn is more annotator-friendly than the standard one inquiring to classify a pixel directly according to our theoretical analysis and user study. Specifically, leveraging foundation models providing useful zero-shot predictions on pseudo labels and superpixels, our method comprises two key techniques: (i) an annotator-friendly design of correction query with the pseudo labels, and (ii) an acquisition function looking ahead label expansions based on the superpixels. Experimental results on PASCAL, Cityscapes, and Kvasir-SEG datasets demonstrate the effectiveness of our ALC framework, outperforming prior methods for active semantic segmentation and label correction. Notably, utilizing our method, we obtained a revised dataset of PASCAL by rectifying errors in 2.6 million pixels in PASCAL dataset.

Modern machine learning applications increasingly demand greater computational resources for training large models. Decentralized training has emerged as an effective means to democratize this technology. However, the potential threats associated with this approach remain inadequately discussed, posing a hurdle to the development of decentralized training infrastructures. This paper aims to initiate discussion towards this end by exploring the robustness of decentralized training from three primary perspectives. Firstly, we articulate our position on establishing robust decentralized training by outlining potential threats and the corresponding countermeasures. Secondly, we illustrate a nascent poisoning attack targeting decentralized training frameworks, easily executable by malicious stages. To mitigate this security threat and ensure efficient training, we propose a robust training framework, integrating a 100% detection strategy and efficient training mechanisms. Finally, we demonstrate the severity of the proposed attack and the effectiveness of our robust training framework. This position paper emphasizes the urgency of exploring the robustness of decentralized training and proposes a feasible solution. The code is available at https://github.com/dcx001016/pipeline_attack.

Large language models (LLM) have emerged as a powerful tool for AI, with the key ability of in-context learning (ICL), where they can perform well on unseen tasks based on a brief series of task examples without necessitating any adjustments to the model parameters. One recent interesting mysterious observation is that models of different scales may have different ICL behaviors: larger models tend to be more sensitive to noise in the test context. This work studies this observation theoretically aiming to improve the understanding of LLM and ICL. We analyze two stylized settings: (1) linear regression with one-layer single-head linear transformers and (2) parity classification with two-layer multiple attention heads transformers (non-linear data and non-linear model). In both settings, we give closed-form optimal solutions and find that smaller models emphasize important hidden features while larger ones cover more hidden features; thus, smaller models are more robust to noise while larger ones are more easily distracted, leading to different ICL behaviors. This sheds light on where transformers pay attention to and how that affects ICL. Preliminary experimental results on large base and chat models provide positive support for our analysis.

In real-world scenarios, the application of reinforcement learning is significantly challenged by complex non-stationarity. Most existing methods attempt to model changes in the environment explicitly, often requiring impractical prior knowledge of environments. In this paper, we propose a new perspective, positing that non-stationarity can propagate and accumulate through complex causal relationships during state transitions, thereby compounding its sophistication and affecting policy learning. We believe that this challenge can be more effectively addressed by implicitly tracing the causal origin of non-stationarity. To this end, we introduce the Causal-Origin REPresentation (COREP) algorithm. COREP primarily employs a guided updating mechanism to learn a stable graph representation for the state, termed as causal-origin representation. By leveraging this representation, the learned policy exhibits impressive resilience to non-stationarity. We supplement our approach with a theoretical analysis grounded in the causal interpretation for non-stationary reinforcement learning, advocating for the validity of the causal-origin representation. Experimental results further demonstrate the superior performance of COREP over existing methods in tackling non-stationarity problems. The code is available at https://github.com/PKU-RL/COREP.

Two models are introduced to study the problem of matching two correlated graphs when some of the nodes are corrupt. In the weak model, a random subset of nodes in one or both graphs can interact randomly with their network. For this model, it is shown that no estimator can correctly recover a positive fraction of the corrupt nodes. Necessary conditions for any estimator to correctly identify and match all the uncorrupt nodes are derived, and it is shown that these conditions are also sufficient for the k-core estimator. In the strong model, an adversarially selected subset of nodes in one or both graphs can interact arbitrarily with their network. For this model, detection of corrupt nodes is impossible. Even so, we show that if only one of the networks is compromised, then under appropriate conditions, the maximum overlap estimator can correctly match a positive fraction of nodes albeit without explicitly identifying them.

Robustness remains a paramount concern in deep reinforcement learning (DRL), with randomized smoothing emerging as a key technique for enhancing this attribute. However, a notable gap exists in the performance of current smoothed DRL agents, often characterized by significantly low clean rewards and weak robustness. In response to this challenge, our study introduces innovative algorithms aimed at training effective smoothed robust DRL agents. We propose S-DQN and S-PPO, novel approaches that demonstrate remarkable improvements in clean rewards, empirical robustness, and robustness guarantee across standard RL benchmarks. Notably, our S-DQN and S-PPO agents not only significantly outperform existing smoothed agents by an average factor of $2.16\times$ under the strongest attack, but also surpass previous robustly-trained agents by an average factor of $2.13\times$. This represents a significant leap forward in the field. Furthermore, we introduce Smoothed Attack, which is $1.89\times$ more effective in decreasing the rewards of smoothed agents than existing adversarial attacks. Our code is available at: [https://github.com/Trustworthy-ML-Lab/Robust_HighUtil_Smoothed_DRL](https://github.com/Trustworthy-ML-Lab/Robust_HighUtil_Smoothed_DRL)

In recent years many methods have been developed to understand the internal workings of neural networks, often by describing the function of individual neurons in the model. However, these methods typically only focus on explaining the very highest activations of a neuron. In this paper we show this is not sufficient, and that the highest activation range is only responsible for a very small percentage of the neuron's causal effect. In addition, inputs causing lower activations are often very different and can't be reliably predicted by only looking at high activations. We propose that neurons should instead be understood as a linear combination of concepts, and develop an efficient method for producing these linear explanations. In addition, we show how to automatically evaluate description quality using simulation, i.e. predicting neuron activations on unseen inputs in vision setting.

Research on the theoretical expressiveness of Graph Neural Networks (GNNs) has developed rapidly, and many methods have been proposed to enhance the expressiveness. However, most methods do not have a uniform expressiveness measure except for a few that strictly follow the $k$-dimensional Weisfeiler-Lehman ($k$-WL) test hierarchy, leading to difficulties in quantitatively comparing their expressiveness. Previous research has attempted to use datasets for measurement, but facing problems with difficulty (any model surpassing 1-WL has nearly 100% accuracy), granularity (models tend to be either 100% correct or near random guess), and scale (only several essentially different graphs involved). To address these limitations, we study the realized expressive power that a practical model instance can achieve using a novel expressiveness dataset, BREC, which poses greater difficulty (with up to 4-WL-indistinguishable graphs), finer granularity (enabling comparison of models between 1-WL and 3-WL), a larger scale (consisting of 800 1-WL-indistinguishable graphs that are non-isomorphic to each other). We synthetically test 23 models with higher-than-1-WL expressiveness on BREC. Our experiment gives the first thorough measurement of the realized expressiveness of those state-of-the-art beyond-1-WL GNN models and reveals the gap between theoretical and realized expressiveness. Dataset and evaluation codes are released at: https://github.com/GraphPKU/BREC.

NLP has recently made exciting progress toward training language models (LMs) with strong scientific problem-solving skills. However, model development has not focused on real-life use-cases of LMs for science, including applications in education that require processing long scientific documents. To address this, we introduce TutorEval and TutorChat. TutorEval is a diverse question-answering benchmark consisting of questions about long chapters from STEM textbooks, written by experts. TutorEval helps measure real-life usability of LMs as scientific assistants, and it is the first benchmark combining long contexts, free-form generation, and multi-disciplinary scientific knowledge. Moreover, we show that fine-tuning base models with existing dialogue datasets leads to poor performance on TutorEval. Therefore, we create TutorChat, a dataset of 80,000 long synthetic dialogues about textbooks. We use TutorChat to fine-tune Llemma models with 7B and 34B parameters. These LM tutors specialized in math have a 32K-token context window, and they excel at TutorEval while performing strongly on GSM8K and MATH. Our datasets build on open-source materials, and we release our models, data, and evaluations publicly.

Graph Neural Networks (GNNs) excel in various graph learning tasks but face computational challenges when applied to large-scale graphs. A promising solution is to remove non-essential edges to reduce the computational overheads in GNN. Previous literature generally falls into two categories: topology-guided and semantic-guided. The former maintains certain graph topological properties yet often underperforms on GNNs. % due to low integration with neural network training. The latter performs well at lower sparsity on GNNs but faces performance collapse at higher sparsity levels. With this in mind, we propose a new research line and concept termed **Graph Sparse Training** **(GST)**, which dynamically manipulates sparsity at the data level. Specifically, GST initially constructs a topology & semantic anchor at a low training cost, followed by performing dynamic sparse training to align the sparse graph with the anchor. We introduce the **Equilibria Sparsification Principle** to guide this process, balancing the preservation of both topological and semantic information. Ultimately, GST produces a sparse graph with maximum topological integrity and no performance degradation. Extensive experiments on 6 datasets and 5 backbones showcase that GST **(I)** identifies subgraphs at higher graph sparsity levels ($1.67\%\sim15.85\%$$\uparrow$) than state-of-the-art sparsification methods, **(II)** preserves more key spectral properties, **(III)** achieves $1.27-3.42\times$ speedup in GNN inference and **(IV)** successfully helps graph adversarial defense and graph lottery tickets.

We study interactive imitation learning, where a learner interactively queries a demonstrating expert for action annotations, aiming to learn a policy that has performance competitive with the expert, using as few annotations as possible. We focus on the general agnostic setting where the expert demonstration policy may not be contained in the policy class used by the learner. We propose a new oracle-efficient algorithm MFTPL-P (abbreviation for Mixed Follow the Perturbed Leader with Poisson perturbations) with provable finite-sample guarantees, under the assumption that the learner is given access to samples from some ``explorative'' distribution over states. Our guarantees hold for any policy class, which is considerably broader than prior state of the art. We further propose Bootstrap-DAgger, a more practical variant that does not require additional sample access.

In-Context Learning (ICL) is the ability of Large Language Models (LLMs) to perform new tasks when conditioned on prompts comprising a few task examples. However, ICL performance can be critically sensitive to the choice of examples. To dynamically select the best examples for every test input, we propose Example Gisting, a novel approach for training example encoders through supervised finetuning with an attention bottleneck between the inputs and outputs. These gist models form the basis for GistScore, a novel metric for scoring and selecting informative examples. Further, we experiment with two variations: (1) finetuning gist models for each dataset and (2) multi-task training a single model on a large collection of datasets. The latter can be used for new tasks out-of-the-box, enabling a training-free ICL pipeline. Evaluations with 21 datasets spanning 9 tasks and 8 diverse LLMs show that our fine-tuned models get state-of-the-art ICL performance with over 20% absolute gain over off-the-shelf retrievers and 5% over the best prior methods. Further, our multi-task model generalizes well to new tasks, datasets, and prompt templates. Selection using this model matches or outperforms prior methods while being three orders of magnitude faster than the strongest training-free baseline.

Discovering reliable and informative relationships among brain regions from functional magnetic resonance imaging (fMRI) signals is essential in phenotypic predictions in neuroscience. Most of the current methods fail to accurately characterize those interactions because they only focus on pairwise connections and overlook the high-order relationships of brain regions. We propose that these high-order relationships should be *maximally informative and minimally redundant* (MIMR). However, identifying such high-order relationships is challenging and under-explored due to the exponential search space and the absence of a tractable objective. In response to this gap, we propose a novel method named HyBRiD, which aims to extract MIMR high-order relationships from fMRI data. HyBRiD employs a Constructor to identify hyperedge structures, and a Weighter to compute a weight for each hyperedge, which avoids searching in exponential space. HyBRiD achieves the MIMR objective through an innovative information bottleneck framework named multi-head drop-bottleneck with theoretical guarantees. Our comprehensive experiments demonstrate the effectiveness of our model. Our model outperforms the state-of-the-art predictive model by an average of 11.2%, regarding the quality of hyperedges measured by CPM, a standard protocol for studying brain connections.

Fourier analysis has been an instrumental tool in the development of signal processing. This leads us to wonder whether this framework could similarly benefit generative modelling. In this paper, we explore this question through the scope of time series diffusion models. More specifically, we analyze whether representing time series in the frequency domain is a useful inductive bias for score-based diffusion models. By starting from the canonical SDE formulation of diffusion in the time domain, we show that a dual diffusion process occurs in the frequency domain with an important nuance: Brownian motions are replaced by what we call mirrored Brownian motions, characterized by mirror symmetries among their components. Building on this insight, we show how to adapt the denoising score matching approach to implement diffusion models in the frequency domain. This results in frequency diffusion models, which we compare to canonical time diffusion models. Our empirical evaluation on real-world datasets, covering various domains like healthcare and finance, shows that frequency diffusion models better capture the training distribution than time diffusion models. We explain this observation by showing that time series from these datasets tend to be more localized in the frequency domain than in the time domain, which makes them easier to model in the former case. All our observations point towards impactful synergies between Fourier analysis and diffusion models.